Dresden 2017 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
BP: Fachverband Biologische Physik
BP 8: Posters - Bioimaging and Spectroscopy
BP 8.11: Poster
Monday, March 20, 2017, 17:30–19:30, P3
GPU-based 3D statistical multi-resolution estimators for image reconstruction — •Stephan Kramer1, Johannes Hagemann2, and Simon Stein3 — 1Fraunhofer ITWM, Kaiserslautern — 2Institut für Röntgenphysik, Universität Göttingen — 3III. Institut für Physik, Universität Göttingen
We extend our previous work [1] on statistical multiresolution estimators (SMRE) to 3D. SMREs are a recent development for the deconvolution of noisy images. Their main feature is the replacement of unphysical regularization parameters by only one which represents the probability that the reconstructed image fulfills the hypothesis test for being the correct one. Their convergence does not degrade with increasing noise. In each iteration step of an SMRE the reconstructed noise is tested whether it obeys a pre-defined distribution on all subsets of the input image. Since the number of possible subsets of an image grow exponentially with its size SMREs are computationally expensive. Reasonable runtimes can only be achieved by a careful parallelization [1] and a proper choice of the image subsets on which the statistical tests are performed. Our implementation is based on our SciPAL library [2] and CUDA. As sample application we discuss preprocessing the time series needed in super-resolution optical fluctuation imaging (SOFI) methods for 3D stacks of images from fluorescence microscopy.
[1] Kramer, S.C., Hagemann, J., Künneke L. and Lebert, J., 2016. SIAM Journal on Scientific Computing, 38(5), pp.C533-C559.
[2] Kramer, S.C. and Hagemann, J., 2015. ACM TOPC, 1(2), p.15.